CN115687894A - Tumble detection system and method based on small sample learning - Google Patents
Tumble detection system and method based on small sample learning Download PDFInfo
- Publication number
- CN115687894A CN115687894A CN202211323361.0A CN202211323361A CN115687894A CN 115687894 A CN115687894 A CN 115687894A CN 202211323361 A CN202211323361 A CN 202211323361A CN 115687894 A CN115687894 A CN 115687894A
- Authority
- CN
- China
- Prior art keywords
- data
- sample
- feature extraction
- network
- csi
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 61
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 45
- 238000000605 extraction Methods 0.000 claims abstract description 44
- 238000012545 processing Methods 0.000 claims abstract description 41
- 238000001914 filtration Methods 0.000 claims abstract description 23
- 230000000694 effects Effects 0.000 claims abstract description 22
- 238000004364 calculation method Methods 0.000 claims abstract description 18
- 230000006399 behavior Effects 0.000 claims abstract description 14
- 230000009471 action Effects 0.000 claims abstract description 9
- 230000002159 abnormal effect Effects 0.000 claims abstract description 8
- 230000008859 change Effects 0.000 claims abstract description 8
- 238000013480 data collection Methods 0.000 claims abstract description 8
- 238000010276 construction Methods 0.000 claims abstract description 4
- 238000007781 pre-processing Methods 0.000 claims abstract description 3
- 230000008569 process Effects 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 11
- 239000013589 supplement Substances 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 10
- 230000007613 environmental effect Effects 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 6
- 238000012886 linear function Methods 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 3
- 230000001131 transforming effect Effects 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 230000008014 freezing Effects 0.000 claims description 2
- 238000007710 freezing Methods 0.000 claims description 2
- 230000008447 perception Effects 0.000 abstract description 5
- 238000013135 deep learning Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 8
- 238000010801 machine learning Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 208000027418 Wounds and injury Diseases 0.000 description 2
- 230000032683 aging Effects 0.000 description 2
- 230000006378 damage Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 208000014674 injury Diseases 0.000 description 2
- 230000000153 supplemental effect Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 230000037081 physical activity Effects 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a fall detection system and method based on small sample learning, which comprises a data collection module, a data processing module and a detection model construction module, wherein a CSI signal which changes due to human activity data in an environment is collected, and each collected CSI signal sample covers continuous signal change within the time of completing an action; preprocessing the collected CSI signals, and performing Hampel filtering processing and discrete wavelet transform filtering processing to remove noise and abnormal values; training two networks, namely a feature extraction network, by using a small sample learning method, and extracting features from the processed CSI signal to map the features to a feature space; and the similarity calculation network is used for calculating the similarity between the extracted features so as to classify the feature data and realize the wireless perception behavior recognition. Compared with the prior art, the invention realizes the wireless perception behavior recognition based on the intelligent algorithm.
Description
Technical Field
The invention relates to the technical fields of signal processing, deep learning, internet of things, wireless sensing and the like, in particular to a fall detection method based on small sample learning.
Background
With the aging development of the population, the aging population increases, and medical care for the elderly becomes increasingly important. The falling is a major factor threatening the life health of the old, the falling of the old is easy to cause serious injury and even endanger life, and the reaction and rescue time after the falling directly influence the injury degree, death risk and later treatment effect of the falling person, so the falling detection of the old is very important.
The traditional fall detection methods can be classified into three categories, one is to use special equipment such as radio frequency radar, ultrasonic waves and laser sensors to collect signal changes caused by falls to detect falls, and the method has high data accuracy, but the deployment cost is high due to the need of deploying the special equipment in the environment. The second type is a method using computer vision, which uses a camera to collect images and processes the image data by analysis to perform fall detection, and the method is seriously affected by ambient light and has privacy problem. The third type is that wearable equipment is used, or sensors are arranged on the body of a testee to collect human body data to detect falling, and the method needs to wear the sensors for a long time and causes inconvenience to the life of the testee. Due to the shortcomings of the conventional methods, a non-invasive, low-cost, easy-to-deploy fall detection scheme is needed. The WiFi signal has the characteristics of non-contact, easiness in acquisition and convenience in deployment, the requirement for fall detection can be met, and the fall detection by using the WiFi signal becomes a novel method with great advantages.
The early WiFi falling detection method uses a traditional machine learning algorithm, after the arrangement equipment collects WiFi signals changed due to human body activities, the signals are subjected to filtering, principal component analysis, fourier transformation and other processing, characteristics are selected from signal samples, and samples to be detected are classified after supervised learning training is carried out by using the machine learning algorithm, so that falling detection is realized. With the development of machine learning algorithms, people find that features which are more abstract and reflect the essence of activities can be extracted in deep learning, compared with the traditional machine learning algorithms, the features of the deep learning algorithms are more flexible and convenient to extract, the extracted abstract features are better in adaptability, and the deep learning starts to be applied to fall detection based on WiFi signals. However, the use of deep learning brings more convenient feature extraction and higher detection accuracy, and also has the problem of large demand for training set data volume. To a certain extent, the deeper the neural network can extract more abstract and more intrinsic characteristics, which means to a certain extent better detection effect and stronger anti-interference capability, but the larger the data required for training the network. On one hand, wiFi signal change data caused by falling needs to be collected by falling of a testee under a real condition, the data collection work is large in workload and high in difficulty, and the testee is easily injured. On the other hand, when the detection environment changes by using the deep learning method, a large amount of data is required for training in the adjustment of the model, which causes that the model is not flexible enough and the environmental adaptability is poor. The invention provides a WiFi falling detection method based on small sample learning by using a meta learning idea and aiming at the problems of the traditional deep learning method, and reduces the falling data volume required in the detection process under the condition of ensuring the precision.
Disclosure of Invention
The invention aims to design a fall detection method based on small sample learning, which combines a deep learning algorithm and an Internet of things technology, and realizes the identification of human fall behaviors by constructing a detection model based on a CSI signal which changes due to collected human activity data in the environment.
The invention is realized by the following technical scheme:
a fall detection system based on small sample learning, the system comprising a data collection module, a data processing module and a detection model construction module, wherein:
the data collection module is used for collecting the CSI signals which are changed due to human activity data in the environment, and each collected CSI signal sample covers the continuous signal change within the time of completing one action;
the data processing module is used for preprocessing the collected CSI signals, and performing Hampel filtering processing and discrete wavelet transform filtering processing to remove noise and abnormal values;
the detection model building module trains two networks by using a small sample learning method, wherein: the characteristic extraction network is used for extracting characteristics from the processed CSI signals and mapping the characteristics to a characteristic space; the similarity calculation network is used for calculating the similarity between the extracted features so as to detect and classify the feature data and identify the falling behavior.
A fall detection method based on small sample learning specifically comprises the following steps:
step 1: collecting data to obtain a CSI signal matrix changed by human activity data in the environment, wherein each collected CSI signal sample covers continuous signal change within the time of completing one action;
step 2: and (3) performing clipping and interpolation processing on the data, wherein: cutting to ensure that the size of a CSI matrix is fixed to be 3 multiplied by 30 multiplied by 200, and performing interpolation processing by using a linear interpolation method, wherein specifically, function values at interpolation points are predicted by linear functions connecting two nearest side points of the interpolation points, so that packet loss data are supplemented;
the linear interpolation processing formula is as follows:
wherein,
supplemental data value, x, for packet loss location 0 、y 0 Is the abscissa/ordinate, x, of a point immediately preceding the position of the patch 1 、y 1 The abscissa/ordinate of the point closest to the supplement position is the abscissa/ordinate of the supplement position, and x is the abscissa of the supplement position;
and step 3: performing data processing, and removing abnormal values of the collected CSI signals caused by environmental interference factors based on Hampel filtering; the Hampel filtering processing comprises the specific process that for each sample function of an input vector, a median of a window consisting of samples and k samples around the samples is obtained, the absolute value of the median is utilized to estimate the standard deviation of each sample to the median, and if the difference between the sample and the median exceeds n standard deviations, the sample is replaced by the median;
the Hampel filtering processing formula is as follows:
wherein X i For the window middle sample, m i Is the median value, σ, of the window i Is the window sample standard deviation, and n is a set threshold;
and 4, step 4: continuing to process data, performing in-band noise filtering based on discrete wavelet transform, performing EMD decomposition on the CSI signal processed in the step 3, transforming the signal to a wavelet domain to obtain a high-frequency IMF component and a low-frequency IMF component of the CSI signal, decomposing the reserved high-frequency IMF component into a high-frequency coefficient and a low-frequency coefficient in the wavelet domain, realizing threshold quantization, namely estimating noise and a threshold of the level in the wavelet transform, adapting the threshold to a lower wavelet level, removing the noise on all the wavelet levels, realizing wavelet threshold denoising, and having no obvious distortion on the signal component; performing wavelet reconstruction, and converting the signals of the wavelet domain back to time domain signals; reconstructing the signal, and reducing to a CSI signal to obtain the CSI signal with high-frequency noise removed;
and 5: constructing a feature extraction model and a detection model, completing feature extraction and similarity calculation, and realizing detection classification of data to be detected so as to identify falling behaviors, wherein the specific process comprises the following steps:
step 5.1: constructing a feature extraction model, namely extracting features from the processed CSI signal by using a feature extraction network to form feature data, mapping the feature data to a feature space, wherein the feature extraction network adopts a multi-layer convolutional neural network structure and extracts relatively abstract features from the feature data;
step 5.2: the method comprises the steps of constructing a detection model, using a similarity calculation network as the detection model to be composed of a plurality of layers of fully-connected networks, training by adopting cross entropy loss, using a network initial parameter matrix to be composed of features extracted from support set samples through a feature extraction network, calculating the similarity between the features extracted from the feature extraction network by using the similarity calculation network, dividing feature data samples into a training set and a support set according to the similarity during detection, wherein the training set comprises different types of collected human activity CSI signal data and is used for training the feature extraction network and the similarity calculation network, the support set comprises a small amount of falling data and activity data, comparing the falling data with the features of the data to be detected after feature extraction, detecting and classifying the data to be detected, and identifying falling behaviors.
Compared with the prior art, the invention can achieve the following beneficial technical effects:
the method can reduce the quantity of data required to be acquired for fall detection, reduce the difficulty and workload of data acquisition, and reduce the overhead of model adjustment required by detection environment change.
Drawings
Fig. 1 is a block diagram of a fall detection system based on small sample learning according to the present invention;
FIG. 2 is a flow chart of the detailed operation of the data processing module;
fig. 3 is an overall flowchart of a fall detection method based on small sample learning according to the present invention;
FIG. 4 is a diagram illustrating a specific process of removing environmental noise for discrete wavelet transform;
FIG. 5 is an exemplary diagram of a feature extraction network training and loss function calculation process;
FIG. 6 is an exemplary diagram of a model training and recognition process.
Detailed Description
The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
The structure, function and operation of the frame according to the present invention will be described in detail below with reference to the accompanying drawings.
Fig. 1 is a block diagram of a WiFi fall detection system based on small sample learning according to the present invention. The system includes a data collection module 100, a data processing module 200, and a detection model construction module 300.
The data collection module 100, by deploying and arranging a signal transmitter and a signal receiver (e.g., an Intel WiFi Link 5300 network card) in the environment, collects a Channel State Information (CSI) signal that changes due to human activity data in the environment using a Linux 802.11n CSI tool. Each collected CSI signal sample covers the duration of the signal change over the time that one action is completed.
The data processing module 200 preprocesses the collected Channel State Information (CSI) data, removes noise and interference, and processes the data into a form convenient for analysis, so as to perform a subsequent analysis processing operation. As shown in fig. 2, a flow chart of the data processing module process is shown. The data processing operation comprises cutting and interpolation processing, and is used for regulating the data size and completing packet loss data. And then, carrying out Hampel filtering processing and discrete wavelet transform filtering processing in sequence for removing data abnormal values and noise. And inputting the processed data into a subsequent model for training and constructing the model.
The detection model building module 300 trains two networks by using a small sample learning method, wherein one feature extraction network is used for extracting features from processed Channel State Information (CSI) signals and mapping signal data to a feature space; and the similarity calculation network is used for calculating the similarity between the extracted features so as to classify the data and realize the wireless perception behavior recognition. The data are divided into a training set and a support set, the training set comprises a large amount of activity data of different types of collected CSI signals, and the activity data are used for training a feature extraction network and a similarity calculation network. The support set comprises a small amount of fall data and other activity data, the fall data and the other activity data are compared with the characteristics of the data to be detected after characteristic extraction, the data to be detected are detected and classified, and fall behaviors are identified. After training using the training set is finished, collecting data of types contained in a small number of support sets as a supplementary training set, and performing Fine-tune Fine adjustment operation on the network to improve detection precision.
Fig. 3 is a flowchart of an overall fall detection method based on small sample learning according to the present invention. Model training and fall identification detection process.
Step 1: collecting data, using two computers provided with Intel WiFi Link 5300 network cards as a signal transmitter and a signal receiver, and using a Linux 802.11n CSI tool to collect CSI signals, wherein the packet sending frequency of the signal transmitter is set to be 20 data packets per second, the length of each action data is 10 seconds (namely 200 data packets), and the signal receiver uses the CSI tool to receive the signals; generating a dat file for each action data sample, and extracting a data matrix of the CSI amplitude signals from the file by using a Matlab corresponding tool, wherein the matrix size of the extracted CSI signals is 3 × 30 × 200 because data come from 30 subcarriers of three antennas;
step 2: cutting and interpolating data to ensure that the size of a CSI matrix is fixed to be 3 multiplied by 30 multiplied by 200, and performing interpolation processing by using a linear interpolation method, wherein specifically, a function value at an interpolation point is predicted by a linear function connecting two nearest side points of the interpolation point, so that packet loss data are supplemented;
the linear interpolation processing formula is as follows:
wherein,
supplemental data value, x, for packet loss location 0 、y 0 Is the transverse/longitudinal direction of the point closest to the previous position of the bag filling positionCoordinate, x 1 、y 1 The abscissa/ordinate of the point closest to the supplement position is the abscissa/ordinate of the supplement position, and x is the abscissa of the supplement position;
and step 3: performing data processing, and removing abnormal values of the collected CSI signals caused by environmental interference factors based on Hampel filtering; the Hampel filtering processing comprises the specific process that for each sample function of an input vector, a median of a window consisting of samples and k samples around the samples is obtained, the absolute value of the median is utilized to estimate the standard deviation of each sample to the median, and if the difference between the sample and the median exceeds n standard deviations, the sample is replaced by the median;
the Hampel filtering processing formula is as follows:
wherein, X i For the window middle sample, m i Is the median value, σ, of the window i Is the window sample standard deviation, and n is a set threshold;
the Hampel filtering can remove abnormal values in the data, and can effectively reduce data abnormality caused by environmental disturbance;
and 4, step 4: the data processing is continued, and the in-band noise filtering is performed based on the discrete wavelet transform, as shown in fig. 4, which is a specific process schematic diagram for removing the environmental noise for the discrete wavelet transform. Performing EMD on the CSI signal processed in the step 3, transforming the signal to a wavelet domain to obtain a high-frequency IMF component and a low-frequency IMF component of the CSI signal, decomposing the reserved high-frequency IMF component into a high-frequency coefficient (detail signal) and a low-frequency coefficient (approximate information) in the wavelet domain, realizing threshold quantization, namely estimating noise and a threshold of the level in wavelet transformation, adapting the threshold to a lower wavelet level, removing the noise on all wavelet levels, realizing wavelet threshold denoising, and having no obvious distortion on the signal component; performing wavelet reconstruction, namely converting the signals of the wavelet domain back to time domain signals; and performing signal reconstruction, namely restoring to the CSI signal to obtain the CSI signal with high-frequency noise removed.
The discrete wavelet transform filtering has a positive effect on removing environmental noise in the wireless perception behavior recognition.
And 5: constructing a feature extraction model and a detection model, completing feature extraction and similarity calculation, and realizing detection classification of data to be detected so as to identify falling behaviors; the specific process is as follows:
step 5.1: constructing a training model, namely extracting features from the processed CSI signal by using a feature extraction network to form feature data, mapping the feature data to a feature space, wherein the feature extraction network adopts a multilayer Convolutional Neural Network (CNN) structure and extracts relatively abstract features from the feature data, a training plan of the feature extraction network adopts triple Loss to train, and three samples, namely an anchor point sample, a forward sample of the same type as the anchor point sample and a reverse sample of different types from the anchor point sample, are selected in each training process; calculating the characteristic distances between the anchor point sample and the forward and reverse samples during training, and training by minimizing the forward characteristic distance and simultaneously maximizing the reverse characteristic distance; the training target is to enable the features of the same type of samples to be distributed close to each other in the feature space and different types to be distributed far from each other in the data features extracted by the feature extraction network. FIG. 5 is a diagram illustrating an example of a process of feature extraction network training and loss function computation.
The model loss function is shown in equation (3).
Wherein f is a feature extraction network, x + For the forward sample, x a For anchor samples, x - For the inverse sample, α is a hyper-parameter representing the model classification boundary value,
is a positive sample x + And anchor point sample x a The distance between the features after feature extraction,
as anchor samples x a And negative sample x + The distance between the features after feature extraction;
step 5.2: and (3) constructing a detection model, namely utilizing a similarity calculation network to calculate the similarity between the features extracted by the feature extraction network, and classifying the samples according to the similarity during detection. The similarity calculation network is composed of a plurality of layers of fully-connected networks and is trained by adopting cross entropy loss, and the network initial parameter matrix is composed of features extracted from each support set sample through a feature extraction network, so that the network training speed and the detection precision are improved. After training is finished, a small amount of data supporting centralized variety actions is used for freezing the lower layer of the network, training the upper layer of the network and Fine-tune adjustment is carried out. FIG. 6 is a diagram illustrating an exemplary model training and recognition process.
According to the invention, a small sample learning method and a meta learning idea are applied to the problem of wireless perception fall detection, so that fall data required to be acquired by model training is greatly reduced, and the data acquisition difficulty and the engineering quantity are reduced; a Fine-tune Fine adjustment method is used in the small sample learning method, so that the detection precision is further improved; when the detection environment is changed, the model can be adjusted by collecting a small amount of new environment data, so that the model can be quickly adapted to the detection in the new environment without retraining.
Claims (5)
1. A fall detection system based on small sample learning, characterized in that the system comprises a data collection module, a data processing module and a detection model construction module, wherein:
the data collection module is used for collecting the CSI signals which are changed due to human activity data in the environment, and each collected CSI signal sample covers the continuous signal change within the time of completing one action;
the data processing module is used for preprocessing the collected CSI signals, and performing Hampel filtering processing and discrete wavelet transform filtering processing to remove noise and abnormal values;
the detection model building module trains two networks by using a small sample learning method, wherein: the characteristic extraction network is used for extracting characteristics from the processed CSI signals and mapping the characteristics to a characteristic space; the similarity calculation network is used for calculating the similarity between the extracted features so as to detect and classify the feature data and identify the falling behavior.
2. A fall detection system based on small sample learning as claimed in claim 1, wherein:
the feature data are divided into a training set and a support set, wherein the training set comprises different types of collected human activity CSI signal data and is used for training a feature extraction network and a similarity calculation network, the support set comprises a small amount of falling data and activity data, the falling data and the activity data are compared with the features of the data to be detected after feature extraction, the data to be detected are detected and classified, and falling behaviors are identified.
3. A fall detection method based on small sample learning is characterized by specifically comprising the following steps:
step 1: collecting data to obtain a CSI signal matrix changed by human activity data in the environment, wherein each collected CSI signal sample covers continuous signal change within the time of completing an action;
step 2: and (3) performing clipping and interpolation processing on the data, wherein: performing cutting processing to ensure that the size of the CSI matrix is fixed to be 3 multiplied by 30 multiplied by 200, and performing interpolation processing by using a linear interpolation method, specifically, predicting a function value at an interpolation point by a linear function connecting two nearest side points of the interpolation point so as to supplement packet loss data;
the linear interpolation processing formula is as follows:
wherein,
supplement for packet loss positionData value, x 0 、y 0 Is the abscissa/ordinate, x, of a point immediately preceding the position of the patch 1 、y 1 The abscissa/ordinate of the point closest to the supplement position is the abscissa/ordinate of the supplement position, and x is the abscissa of the supplement position;
and step 3: performing data processing, and removing abnormal values of the collected CSI signals caused by environmental interference factors based on Hampel filtering; the Hampel filtering processing comprises the specific process that for each sample function of an input vector, a median of a window consisting of samples and k samples around the samples is obtained, the absolute value of the median is utilized to estimate the standard deviation of each sample to the median, and if the difference between the sample and the median exceeds n standard deviations, the sample is replaced by the median;
the Hampel filtering processing formula is as follows:
wherein, X i For the window middle sample, m i Is the median value, σ, of the window i Is the window sample standard deviation, and n is a set threshold;
and 4, step 4: continuing to process data, performing in-band noise filtering based on discrete wavelet transform, performing EMD decomposition on the CSI signal processed in the step 3, transforming the signal to a wavelet domain to obtain a high-frequency IMF component and a low-frequency IMF component of the CSI signal, decomposing the reserved high-frequency IMF component into a high-frequency coefficient and a low-frequency coefficient in the wavelet domain, realizing threshold quantization, namely estimating noise and a threshold of the level in the wavelet transform, adapting the threshold to a lower wavelet level, removing the noise on all the wavelet levels, realizing wavelet threshold denoising, and having no obvious distortion on the signal component; performing wavelet reconstruction, and converting the signals of the wavelet domain back to time domain signals; reconstructing the signal, and reducing to a CSI signal to obtain the CSI signal with high-frequency noise removed;
and 5: constructing a feature extraction model and a detection model, completing feature extraction and similarity calculation, and realizing detection classification of data to be detected so as to identify falling behaviors, wherein the specific process comprises the following steps:
step 5.1: constructing a feature extraction model, namely extracting features from the processed CSI signals by using a feature extraction network to form feature data, mapping the feature data to a feature space, and extracting relatively abstract features from the feature data by using a multi-layer convolution neural network structure of the feature extraction network;
step 5.2: the method comprises the steps of constructing a detection model, namely training a similarity calculation network, wherein an initial parameter matrix is composed of features extracted from support set samples through a feature extraction network, the similarity between the features extracted from the feature extraction network is obtained through the similarity calculation network, the feature data samples are divided into a training set and a support set according to the similarity during detection, the training set comprises different types of collected human activity CSI signal data and is used for training the feature extraction network and the similarity calculation network, the support set comprises a small amount of falling data and activity data, the falling data are compared with the features of the data to be detected after feature extraction, the data to be detected are detected and classified, and falling behaviors are identified.
4. The fall detection method based on small sample learning as claimed in claim 3, wherein the feature extraction network is trained by triple Loss, and three samples are selected in each training process, namely an anchor sample, a forward sample of the same type as the anchor sample, and a backward sample of a different type from the anchor sample; calculating the characteristic distances between the anchor point sample and the forward and reverse samples during training, and training by minimizing the forward characteristic distance and simultaneously maximizing the reverse characteristic distance;
the model loss function is shown in equation (3):
wherein f is a feature extraction network, x + Is a forward sample, x a For anchor samples, x - For the inverse sample, α is a hyper-parameter representing the model classification boundary value,
is a positive sample x + And anchor point sample x a The distance between the features after feature extraction,
as anchor samples x a And negative sample x + The distance between the features after feature extraction.
5. A fall detection method based on small sample learning as claimed in claim 3, wherein after training using the training set is finished, the data in the support set is collected as a supplementary training set, network low-level freezing and network high-level training are realized by using the data in the support set, and Fine-tune operation is performed on the network to improve detection accuracy.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211323361.0A CN115687894A (en) | 2022-10-27 | 2022-10-27 | Tumble detection system and method based on small sample learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211323361.0A CN115687894A (en) | 2022-10-27 | 2022-10-27 | Tumble detection system and method based on small sample learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115687894A true CN115687894A (en) | 2023-02-03 |
Family
ID=85099497
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211323361.0A Pending CN115687894A (en) | 2022-10-27 | 2022-10-27 | Tumble detection system and method based on small sample learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115687894A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116821778A (en) * | 2023-08-30 | 2023-09-29 | 之江实验室 | Gait recognition method and device based on WIFI signals and readable storage medium |
| CN117496243A (en) * | 2023-11-06 | 2024-02-02 | 南宁师范大学 | Small sample classification method and system based on contrast learning |
-
2022
- 2022-10-27 CN CN202211323361.0A patent/CN115687894A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116821778A (en) * | 2023-08-30 | 2023-09-29 | 之江实验室 | Gait recognition method and device based on WIFI signals and readable storage medium |
| CN116821778B (en) * | 2023-08-30 | 2024-01-09 | 之江实验室 | Gait recognition method and device based on WIFI signals and readable storage medium |
| CN117496243A (en) * | 2023-11-06 | 2024-02-02 | 南宁师范大学 | Small sample classification method and system based on contrast learning |
| CN117496243B (en) * | 2023-11-06 | 2024-05-31 | 南宁师范大学 | Small sample classification method and system based on contrast learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107844755B (en) | Electroencephalogram characteristic extraction and classification method combining DAE and CNN | |
| CN115687894A (en) | Tumble detection system and method based on small sample learning | |
| CN105919584B (en) | Heart rate method of estimation and device for wearable heart rate monitor apparatus | |
| CN108416367B (en) | Sleep staging method based on multi-sensor data decision-level fusion | |
| CN110151203B (en) | Fatigue driving identification method based on multistage avalanche convolution recursive network EEG analysis | |
| CN111134662B (en) | Electrocardio abnormal signal identification method and device based on transfer learning and confidence degree selection | |
| CN103190904B (en) | Electroencephalogram classification detection device based on lacuna characteristics | |
| CN103610466A (en) | EMG fall identification method based on EMD permutation entropy | |
| CN114052744B (en) | Electrocardiosignal classification method based on impulse neural network | |
| CN111178331A (en) | Radar image recognition system, method, apparatus, and computer-readable storage medium | |
| CN111387975A (en) | Electroencephalogram signal identification method based on machine learning | |
| CN113010013A (en) | Wasserstein distance-based motor imagery electroencephalogram migration learning method | |
| Al-Assaf | Surface myoelectric signal analysis: dynamic approaches for change detection and classification | |
| CN113935373A (en) | Human body action recognition method based on phase information and signal intensity | |
| CN112017155A (en) | Method, device and system for measuring health sign data and storage medium | |
| CN114999643A (en) | WiFi-based intelligent monitoring method for old people | |
| Moshou et al. | Wavelets and self-organising maps in electromyogram (EMG) analysis | |
| CN114595725A (en) | Electroencephalogram signal classification method based on addition network and supervised contrast learning | |
| CN116595434A (en) | Lie detection method based on dimension and classification algorithm | |
| CN113397561B (en) | Electroencephalogram detection system based on S transformation and Bayes convolutional neural network and working method and device thereof | |
| CN112883355B (en) | Non-contact user identity authentication method based on RFID and convolutional neural network | |
| CN116264012A (en) | Contactless falling direction detection method for short message alarm | |
| CN114499712A (en) | Gesture recognition method, device and storage medium | |
| CN102222217A (en) | Human body action pyroelectric device and identification method thereof | |
| Lu et al. | Uterine electromyography signal feature extraction and classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |